基于卷积神经网络与可见近红外光谱的土壤含氮量检测

Detection of Soil Nitrogen Content Based on Convolution Neural Network and Visible-near Infrared Spectroscopy

为实现土壤氮含量的快速检测,提出一种基于卷积神经网络(CNN)与可见近红外光谱的土壤氮含量检测方法。采用批量正则化及dropout技术提升模型性能,降低过拟合。实验中,对模型的训练次数及dropout丢弃比例进行了对比选择,并将结果与传统的PLSR及BP神经网络进行对比。结果表明,CNN模型的预测精度与泛化能力明显优于传统模型,测试集的决定系数分别比PLSR和BP神经网络高0.2731和0.1686,这表明CNN模型能够实现对土壤可见近红外光谱数据的特征提取,进而实现对土壤氮含量的快速检测,为后续的土壤养分快速检测仪器开发提供了基础。

In order to realize the rapid detection of soil nitrogen content,a method based on convolutional neural network(CNN)and visible near infrared spectroscopy(NIRS)was proposed.Batch normalization and dropout technology were used to improve the model performance and reduce over fitting.In the experiment,the training times and dropout discarding ratio of the model were compared and selected,and the results were compared with the traditional PLSR and BP neural network.The results showed that the prediction accuracy and generalization ability of CNN model were significantly better than the traditional model,and the determination coefficient of test set was 0.2731 and 0.1686higher than that of PLSR and BP neural network,respectively,which showed that CNN model could realize the feature extraction of soil visible and near-infrared spectrum data,and then realize the rapid detection of soil nitrogen content,which provided the basis for the development of soil nutrient rapid detection instrument.